ICML 2026 论文列表

Off-policy learning methods seek to derive an optimal policy directly from a fixed dataset of prior interactions. This objective presents significant challenges, primarily due to the inherent distributional shift and value function overestimation bias. These issues become even more noticeable in zero-shot reinforcement learning, where an agent trained on reward-free data must adapt to new tasks at test time without additional training. In this work, we address the off-policy problem in a zero-shot setting by discovering a theoretical connection of successor measures to stationary density ratios. Using this insight, our algorithm can infer optimal importance sampling ratios, effectively performing a stationary distribution correction with an optimal policy for any task on the fly. We benchmark our method in motion tracking tasks on SMPL Humanoid, continuous control on ExoRL, and for the long-horizon OGBench tasks. Our technique seamlessly integrates into forward-backward representation frameworks and enables fast-adaptation to new tasks in a training-free regime. More broadly, this work bridges off-policy learning and zero-shot adaptation, offering benefits to both research areas.

We present WorldMirror, a unified feed-forward model for comprehensive 3D geometric prediction tasks. Unlike existing methods constrained to image-only inputs or customized for a specific task, our framework flexibly integrates diverse geometric priors, including camera poses, intrinsics, and depth maps, while simultaneously generating multiple 3D representations: dense point clouds, multi-view depth maps, camera parameters, surface normals, and 3D Gaussians. Remarkably, prior injection yields universal gains across all tasks, suggesting that input flexibility and multi-task prediction are mutually reinforcing. WorldMirror achieves state-of-the-art performance across diverse benchmarks from camera, point map, depth, and surface normal estimation to novel view synthesis, while maintaining the efficiency of feed-forward inference. Code and model weights are publicly available at https://github.com/Tencent-Hunyuan/HunyuanWorld-Mirror.

Large Language Models (LLMs) often struggle with the precise logic and schema alignment required for complex Text-to-SQL tasks. While current methods rely heavily on static prompting, they lack the ability to dynamically adapt and self-correct through environmental interaction. To bridge this gap, we propose **MARS-SQL**, a trainable multi-agent framework for Text-to-SQL. Rather than introducing a new standalone SQL primitive, **MARS-SQL** makes an agentic workflow trainable by decomposing the problem into three specialized roles: schema grounding, query generation, and solution validation. Central to our approach is a generation agent trained via a multi-turn RL policy within a ReAct-style loop. The agent learns to iteratively reason, execute intermediate SQL actions on a live database, and refine its strategy based on execution feedback. To improve robustness, we further introduce a validation mechanism that treats solution selection as a generative modeling task, identifying the optimal interaction trajectory through next-token prediction probabilities. Empirical evaluations demonstrate the effectiveness of coupling interactive learning with trajectory ranking. **MARS-SQL** achieves state-of-the-art performance, recording an execution accuracy of 77.84\% on the BIRD development dataset and 89.75\% on the Spider test dataset, while also transferring strongly to out-of-domain benchmarks. Code is available at https://github.com/YangHaolin0526/MARS-SQL.

Antimicrobial peptide discovery is challenged by the astronomical size of peptide space and the relative scarcity of active peptides. While generative models provide latent maps of this space, they typically ignore decoder-induced geometry and rely on flat Euclidean metrics, making exploration distorted and inefficient. Existing manifold-based approaches assume fixed intrinsic dimensionality, which fails for real peptide data. We introduce **PepCompass**, a geometry-aware framework based on a **Union of $\kappa$-Stable Riemannian Manifolds** that captures local decoder geometry while maintaining computational stability. PepCompass performs global interpolation via **Potential-minimizing Geodesic Search (PoGS)** to bias discovery toward promising seeds and enables local exploration through **Second-Order Riemannian Brownian Efficient Sampling** and **Mutation Enumeration in Tangent Space**, which together form **Local Enumeration Bayesian Optimization (LE-BO)**. PepCompass achieves a 100% *in-vitro* validation rate: PoGS identifies four novel seeds and LE-BO optimizes them into 25 highly active, broad-spectrum peptides, demonstrating that geometry-informed exploration is a powerful paradigm for antimicrobial peptide design.

The two-sample testing problem, a fundamental task in statistics and machine learning, seeks to determine whether two sets of samples, drawn from underlying distributions $p$ and $q$, are in fact identically distributed (i.e.~whether $p=q$). A popular and intuitive approach is the classifier two-sample test (C2ST), where a classifier is trained to distinguish between samples from $p$ and $q$. Yet despite simplicity of the C2ST, its reliability hinges on access to a near-Bayes-optimal classifier, a requirement that is rarely met and difficult to verify. This raises a major open question: can a weak classifier still be useful for two-sample testing? We show that the answer is a definitive yes. Building on the work of Hu & Lei (2024), we analyze two conformal variants of the C2ST that convert the scores from any trained classifier---even if weak, biased, or overfit---into exact, finite-sample p-values. We establish two key theoretical properties of the conformal C2ST: (i) finite-sample Type-I error control, and (ii) non-trivial power that degrades gently in tandem with the error of the trained classifier. The upshot is that even poorly performing classifiers can yield powerful and reliable two-sample tests. This general framework finds a powerful application in Bayesian inference, particularly for validating Neural Posterior Estimation (NPE) models, where the task of comparing a learned posterior approximation $q(\theta \mid y)$ to the true posterior $p(\theta \mid y)$ can be framed as a two-sample test. Empirically, the Conformal C2ST outperforms classical discriminative tests across a wide range of benchmarks for this task. Our results establish the conformal C2ST as a practical, theoretically grounded diagnostic tool.

Agents that can reason and plan in the real world must be able to predict the consequences of their actions. World models possess this capability but require action annotations that can be complex to obtain at scale. Latent action models address this issue by learning an action space from videos alone. Our work studies the training of latent action world models on in-the-wild videos, expanding the scope of existing works that focus on simple robotics simulations, video games, or manipulation data. While diverse videos enable modeling richer actions, they introduce challenges of environmental noise and lack of a common embodiment across videos. To address these, we carefully study the design and evaluation of latent actions. We find that constrained continuous latent actions are better suited for complex in-the-wild videos, compared to vector quantization. For example, actions specific to in-the-wild videos such as humans entering the room, can be modeled and then transferred across videos. However, in the absence of a common embodiment, learned latent actions are localized in space, relative to the camera. Nonetheless, we are able to train a controller that maps known actions to latent ones, allowing us to use latent actions as a universal interface to solve planning tasks on par with action-conditioned baselines.

We consider the privacy amplification properties of a sampling scheme in which a user’s data is used in $k$ steps chosen randomly and uniformly from a sequence (or set) of $t$ steps. This sampling scheme has been recently applied in the context of differentially private optimization [Chua et al., 2024a, Choquette-Choo et al., 2025] and communication-efficient high-dimensional private aggregation [Asi et al., 2025], where it was shown to have utility advantages over the standard Poisson sampling. Theoretical analyses of this sampling scheme [Feldman and Shenfeld, 2025, Dong et al., 2025] lead to bounds that are close to those of Poisson sampling, yet still have two significant shortcomings. First, in many practical settings, the resulting privacy parameters are not tight due to the approximation steps in the analysis. Second, the computed parameters are either the hockey stick or Rényi divergence, both of which introduce overheads when used in privacy loss accounting. In this work, we demonstrate that the privacy loss distribution (PLD) of random allocation applied to any differentially private algorithm can be computed efficiently. When applied to the Gaussian mechanism, our results demonstrate that the privacy-utility trade-off for random allocation is at least as good as that of Poisson subsampling. In particular, random allocation is better suited for training via DP-SGD. To support these computations, our work develops new tools for general privacy loss accounting based on a notion of PLD realization. This notion allows us to extend accurate privacy loss accounting to subsampling which previously required manual noise-mechanism-specific analysis.

The robustness of synchronization is typically characterized by scalar, per-node stability indices whose dependence on topology is studied via network science or graph neural networks (GNNs). We propose a novel upstream task, stability landscapes, which provide deeper insights into synchronization behavior and from which many such scalar indices can be derived. Crucially, we pioneer a graph-to-image prediction paradigm: learning image-like landscapes as per-node targets directly from graph topology, a formulation we are not aware of having been established elsewhere in the literature. To support this task, we release two datasets of 10,000 graphs each at 20 and 100 nodes with per-node landscape labels, based on a conceptual oscillator model, capturing power grid synchronization behavior. A GNN encodes topology and a CNN decoder renders per-node images, learned end-to-end with good in-distribution accuracy, generalizing across graph sizes and to realistic power grid topologies. This demonstrates that stability landscapes, while beyond the reach of conventional network science, are learnable from topology and open new avenues for moving beyond scalar stability indices in biology, neuroscience, and power grids.

Process rewards have been widely used in deep reinforcement learning to improve training efficiency, reduce variance, and prevent reward hacking. In LLM reasoning, existing works also explore various solutions for learning effective process reward models (PRM) with or without the help of an expert policy. However, existing methods either rely on strong assumptions about the expert policies (e.g., requiring their reward functions) or suffer intrinsic limitations (e.g., entropy collapse), resulting in weak PRMs or limited generalizability. In this paper, we introduce rePIRL, an inverse RL-inspired framework that learns effective PRMs with minimal assumptions about expert policies. Specifically, we design a dual learning process that updates the policy and the PRM interchangeably. Our learning algorithm has customized techniques to address the challenges of scaling traditional inverse RL to LLMs. We theoretically show that our proposed learning framework can unify both online and offline PRM learning methods, justifying that rePIRL can learn PRMs with minimal assumptions. Empirical evaluations on standardized math and coding reasoning datasets demonstrate the effectiveness of rePIRL over existing methods. We further show the application of our trained PRM in test-time training, test-time scaling, and providing an early signal for training hard problems. Finally, we validate our training recipe and key design choices via a detailed ablation study.

Infrared and visible video fusion is essential for achieving comprehensive perception in dynamic scenes. However, maintaining temporal consistency remains a formidable challenge. Conventional methods relying on optical flow often suffer from geometric rigidity and ghosting artifacts. Moreover, standard diffusion-based fusion models typically operate in a frame-by-frame manner; when extended to autoregressive settings, they lack intrinsic temporal constraints and are prone to severe error accumulation and drifting, where minor artifacts amplify over time. To address these limitations, we propose a drift-resilient video fusion method that reformulates the task as history-conditioned motion generation. We introduce Stabilized History Guidance and Soft Temporal Anchoring to reframe temporal consistency as spectral filtering, implicitly aggregating motion dynamics without rigid alignment. Furthermore, our Decoupled Structure-Motion Adaptation strategy bridges pre-trained priors and structural constraints via two-stage training and latent refinement. Extensive experiments demonstrate that our method achieves state-of-the-art performance in both fusion quality and temporal stability.

Many learning problems require uncovering a hidden ordering that reveals structure in unordered data, such as monotonicity in sorting or spatial continuity in jigsaw reconstruction. In these settings, permutations can be learned as latent operators by optimizing objectives defined directly on the reordered output, often without access to ground-truth orderings. Differentiable relaxations such as Gumbel–Sinkhorn make this approach practical by approximating permutation matrices with doubly stochastic matrices. However, learning from structure without supervision induces a non-uniform uncertainty: some assignments become confident early, while others remain ambiguous. Existing methods control this process using a single global temperature, forcing all assignments to sharpen or diffuse simultaneously and leading to instability at scale. We introduce an entropy-adaptive formulation of Gumbel–Sinkhorn that locally modulates temperature based on assignment uncertainty. This allows confident assignments to discretize early while preserving exploration where uncertainty remains. Across sorting and jigsaw reconstruction tasks and in routing-style settings, adaptive entropy control improves training stability and final permutation quality relative to fixed-temperature baselines, particularly as problem size and assignment ambiguity increase.

Modern computer-use agents (CUA) must perceive a screen as a structured state, what elements are visible, where they are, and what text they contain, before they can reliably ground instructions and act. Yet, most available grounding datasets provide sparse supervision, with *insufficient* and *low-diversity* labels that annotate only a small subset of task-relevant elements per screen, which limits both coverage and generalization; moreover, practical deployment requires efficiency to enable low-latency, on-device use. We introduce **ScreenParse**, a large-scale dataset for *complete* screen parsing, with dense annotations of all visible UI elements (boxes, 55-class types, and text) across 771K web screenshots (21M elements). ScreenParse is generated by **Webshot**, an automated, scalable pipeline that renders diverse urls, extracts annotations and applies VLM-based relabeling and quality filtering. Using ScreenParse, we train **ScreenVLM**, a compact, 316M-parameter vision language model (VLM) that decodes a compact ScreenTag markup representation with a structure-aware loss that upweights structure-critical tokens. ScreenVLM substantially outperforms much larger foundation VLMs on dense parsing (e.g., 0.592 vs. 0.294 PageIoU on ScreenParse) and shows strong transfer to public benchmarks. Moreover, finetuning foundation VLMs on ScreenParse consistently improves their grounding performance, suggesting that dense screen supervision provides transferable structural priors for UI understanding. The dataset, model and code will be made publicly available.

Few shot segmentation with vision foundation models relies on high quality prompts to segment unseen categories from limited support annotations. Existing prompt construction methods depend on test-time adaptation, fixed heuristic sampling, or monolithic reinforcement learning, making it difficult to balance semantic consistency, spatial coverage, and prompt credit assignment. To address these limitations, PromptPilot is proposed as a hierarchical multi-agent reinforcement learning framework for point prompt optimization with frozen DINOv2 and SAM. Prompt construction is formulated as sequential decision making, where feature and physical agents propose complementary prompt modifications, and a manager agent selects actions using SAM feedback and local marginal contribution. PromptPilot functions as an inference-time optimization strategy without parameter updates. Extensive experiments demonstrate that the proposed game-theoretic approach improves segmentation performance and generalization, offering a principled solution for automated prompt engineering. The code is available at [https://github.com/L-AILab/PromptPilot](https://github.com/L-AILab/PromptPilot).

We introduce a predictive model that estimates the pre-training loss of large models from model size ($N$), batch size ($B$) and number of weight updates ($K$). This is the first loss prediction model that can handle changing batch size. The model outperforms Chinchilla's loss model, a model of the test loss using the batch size and number of tokens, in terms of projecting the loss at extrapolated compute budgets (up to 1000 folds). A natural use of the model is to find optimal $N,B,K$ configurations under explicit and compound resource constraints like time, memory and compute. In our experiments, the model-selected configurations are close to ground-truth optimal. Our work advocates for loss prediction as a better alternative to heuristic-based laws, which are growing in complexity. The implementation is available on \href{https://github.com/chuningxdy/Noisy-Quadratic-System}{GitHub}.

Large Language Models (LLMs) can propose natural-language rules, circumventing the reliance on a predefined predicate space in traditional rule learning. However, existing LLM-based methods often neglect the global interactions among rules, and the potential of using fine-grained rule importance scores to calibrate neuro-symbolic reasoning remains underexplored. To address this gap, we introduce RLIE, a framework that integrates LLMs with probabilistic modeling to learn weighted rule sets in four stages: (1) Rule generation: proposing and filtering candidate rules via LLMs; (2) Logistic regression: learning sparse, calibrated weights for global rule selection; (3) Iterative refinement: revising the rule set with error-driven hard examples; and (4) Evaluation: validating the learned system via comparative inference paradigms. Across multiple real-world datasets and LLM backbones, our learned weighted rules achieve superior stability and accuracy, whereas rule-injection prompting yields mixed results and often degrades performance. These results suggest LLMs excel at semantic rule discovery but are less reliable at controlled probabilistic aggregation. Our findings highlight both the promise and the limits of LLMs for inductive reasoning, motivating a principled integration with classic probabilistic rule combination for reliable neuro-symbolic reasoning.

Attention is the critical component of a transformer. Yet the quadratic computational complexity of vanilla full attention in the input size and the inability of its linear attention variant to focus have been challenges for computer vision tasks. We provide a mathematical definition of generalized attention and formulate both vanilla softmax attention and linear attention within the general framework. We prove that generalized attention disperses, that is, as the number of keys tends to infinity, the query assigns equal weights to all keys. Motivated by the dispersion property and recent development of Mamba form of attention, we design Scalable and Efficient Mamba like Attention (SEMA) which utilizes token localization to avoid dispersion and maintain focusing, complemented by theoretically consistent arithmetic averaging to capture global aspect of attention. We support our approach on Imagenet-1k where classification results show that SEMA is a scalable and effective alternative beyond linear attention, outperforming recent vision Mamba models on increasingly larger scales of images at similar model parameter sizes. Source code can be found at: https://github.com/nhatthanhtran/SEMA.

Long-tailed recognition fundamentally suffers from optimizer blindness where the optimization process mistakenly conflates the magnitude of gradient accumulation with the scarcity of semantic information. Existing strategies relying on static frequency-based priors fail to correct this bias and result in state blindness regarding supervision and micro-level blindness regarding parameter updates. To address these limitations, we propose the AES framework to establish a dynamic and state-aware correction system across the entire learning lifecycle. We specifically introduce Adaptive Residual Supervision loss to act as a real-time reality check for supervision completeness via precision shielding. We also propose Entropy-aware PCGrad to resolve parameter-level conflicts by quantifying task specificity through gradient entropy. Additionally, we devise Sample-level Conflict Arbitrated Fusion to serve as a dynamic inference arbiter that routes predictions based on instance difficulty. Extensive experiments on CIFAR-100-LT, ImageNet-LT, and iNaturalist 2018 demonstrate that our method consistently achieves state-of-the-art performance by effectively balancing head-class stability and tail-class discrimination. Code is available at [here](https://github.com/wangff0101/AES)

Masked diffusion large language models (dLLMs) are a promising alternative to autoregressive generation. While reinforcement learning (RL) algorithms have been adapted to be compatible with dLLMs for fine-tuning them, their reliance on the computation of the marginal likelihood to evaluate policy objectives is intractable. To overcome this, we exploit a dynamical relation between the unmasking posterior of the base model and that which targets the reward-tilted distribution to derive Discrete Tilt Matching (DTM), an algorithm that avoids intractable likelihood evaluation entirely. DTM can be phrased as a cross-entropy loss that only requires forward evaluation of rewards and whose variance can be adaptively controlled, improving training stability. We motivate DTM on maze planning tasks, and show that fine-tuning LLaDA-8B-Instruct with DTM achieves higher accuracy at lower compute costs than prior RL-based fine-tuning methods across the Sudoku, Countdown, and MATH500 benchmarks.

Despite the success of LVLMs, general optimization objectives (e.g., standard MLE) fail to constrain visual trajectories, leading to language bias and hallucination. To mitigate this, current methods introduce geometric priors from visual experts as additional supervision. However, we observe that such supervision is typically suboptimal: *it is biased toward geometric precision and offers limited reasoning utility*. To bridge this gap, we propose Perceptual Flow Network (PFlowNet), which eschews rigid alignment with the expert priors and achieves interpretable yet more effective visual reasoning. Specifically, PFlowNet decouples perception from reasoning to establish a self-conditioned generation process. Based on this, it integrates *multi-dimensional rewards* with *vicinal geometric shaping* via variational reinforcement learning, thereby facilitating reasoning-oriented perceptual behaviors while preserving visual reliability. PFlowNet delivers a provable performance guarantee and competitive empirical results, particularly setting new SOTA records on V* Bench (90.6%) and MME-RealWorld-lite (67.0%).